Forward sweeps predict human planning via model-based roll-outs but not successor-representations

Navigating rodents show “forward sweeps” - sequential reactivation of non-local neural representations of future states starting from their current position. While direct evidence remains scant, it is widely thought that such forward sweeps support model-based (MB) evaluation of actions. However, a key challenge in testing this hypothesis is the absence of experimental paradigms that allow for precise measurement of MB action evaluation. Notably, typical assessments of MB evaluation, i.e. flexibility following changes to rewards (reward revaluation), conflate reliance on full MB evaluation with a simpler approximation strategy, the successor representation (SR). Whereas full MB evaluates actions by sequentially rolling out a one-step model of state transitions, SR relies on cached temporally-flat predictions of states over multiple future time-steps, learned directly from experience. We tested the hypothesis that neural forward sweeps specifically implement forward roll-outs, and thus MB, but not SR evaluation. We designed a novel, highly-powered, behavioural task for MEG, capable of disambiguating behavioural evidence for MB and SR evaluation as well identifying forward sweeps through the task transition model. We found participants’ (N = 30) choices reflected evidence for use of both MB and SR evaluation strategies. We also found that participant’s response times were modulated by MB planning depth, and that this effect was greater in participants whose choices more reflected full MB planning, linking the behavioural measure of MB choice to sequential rollouts. Finally, we found that, across participants, neural evidence for forward sweeps, assessed by decoded sequential reactivations of upcoming task states, was related to behavioural evidence for use of MB, but not SR, choice evaluation. This provides the first direct evidence that forward sweeps support MB evaluation via roll-outs.